ENHANCED LIKELIHOOD COMPUTATION USING REGRESSION

Abstract

In a rank based large vocabulary continuous speech recognition system [1], the correct leaf is expected to occupy the top rank positions. An increase in the number of times the correct leaf occurs in the top rank positions translates to an increase in word accuracy. In order to achieve low error rates, we need to discriminate the most confusable incorrect leaves from the correct leaf by lowering their ranks. Therefore, the goal here is to increase the likelihood of the correct leaf of a frame, while decreasing the likelihoods of the confusable leaves. In order to do this, we use the auxiliary information from the prediction of the neighboring frames to augment the likelihood computation of the current frame. We then use the residual errors in the predictions of neighboring frames to discriminate between the correct (best) and incorrect leaves of a given frame. In this paper, we present a new algorithm that incorporates prediction error likelihoods into the overall likelihood computation to improve the rank position of the correct leaf. Experimental results on the Wall Street Journal task and an in-house large vocabulary continuous speech recognition task show a relative accuracy improvements in speaker-independent performance of 10%.